Facebook M – The Anti-Turing Test
medium.com
medium.com
This sounds like a 10 page side story in a Neil Stephenson Book. I love it.
The author is also a relatively prolific HN poster (https://news.ycombinator.com/user?id=cstross).
The spam tries to act like a perfectly normal message as long as it is talking to the spam filter, and as soon as it thinks it is talking to a real person, it shows its spam message. The spam filter tries to impersonate the recipient as best as it can (in 3D video), meanwhile trying to figure out whether the message is spam.
The spam filters are unfortunately humstrung by the fact that they can only become close to real conscious AI and not further, because taking them all the way there would mean you'd be exposing a conscious being to spam all its life, which would be torture and thus criminal. Spammers don't care.
IIRC, this is just a side anecdote in some paragraphs somewhere, but I love it.
I don't know anybody like Egan, he takes the science part to such extremes... My favourite SF author is Iain M Banks, but you're probably already familiar with him.
Whether Comcast would want to make it that easy for people to cancel is entirely a different issue though :)
The implication to me is that the chat is with a human, who is using an AI tool with the intention of training that tool. What better way to train a new service than to launch it, then answer all the weird, unexpected questions with humans? Gradually more of the questions get answered, the AI gets better trained, and the human-AI becomes increasing more AI.
Further, as the AI gets better, the human working with it has to do less, so they can roll out the service to more users without requiring more staff. Perhaps eventually, no human is needed.
Bot reads "What's the temperature like near me?"
Person calls "$get user-local-temperature"
API responds "{temperature:{f:77},{c:99}}"
Human writes "It's 77 degrees outside!"
Bot reads "What's it like outside?"
Person calls "$get user-local-weather"
API responds "{weather:{now:Sunny},{today:Cold}}"
Human writes "It's sunny now, but will be cold later today."
There's no serious corpus yet for that -- if this is real, it is important work.
It doesn't need to see "What|WP 's|VBZ it|PRP like|IN outside|IN ?|."
Just "What's it like outside?", and know that it will always respond to that with the same call.
Warning: This is mere conjecture.
The data around how users interact will also be important:
Do they prefer a back and forth conversation, or do they want to say everything in one go.
Do they want to start a conversation, drop it, come back to it several hours later, or do they like completing it in one go.
How do they handle switching back and forth between different contexts, if certain requests take time, or do users not switch context.
What data are they happy to share, and what are they not.
What are the typical response times that a user considers acceptable 5 seconds, 1 minute, 5 minutes, 60 minutes? Does it vary depending on scenario.
Is there particular services or information that there is a trend towards, for example local search requests, research/information, particular types of purchase etc.
We've just spent 6 months going through a very similar process to this, which has helped drive the development of our Converse platform, which allows people to build semi or fully automated conversational messaging services, so this is fairly closely related.
(From our point of view, the NLP data we gathered was useful, but it wasn't the most important part)
The presence of a delay does not mean there is no A.I. there. Not everything is as fast as Google search, for instance IBM Watson would think about a problem for a few seconds, which is fine because it only needs to be about as fast as a human.
It's also strikingly similar to the original "mechanical turk".
That's what the M stands for.
http://recode.net/2015/11/03/facebooks-virtual-assistant-m-i...
Clearly, the author didn't even do the most basic fact checking. Since, Facebook clearly told everyone that M was going to be AI that was assisted by humans.
It's literally in the announcement post: https://www.facebook.com/Davemarcus/posts/10156070660595195
> "It's powered by artificial intelligence that's trained and supervised by people."
The AI can still do the requirements/information gathering and leaving the job for a person to do. For common issues (support triage, customer service issues) even this still has significant value.
Equally from a training point of view, it may simply be the person correcting/confirming the AI is right, and leaving it to get on with the process, rather than fallover to the human completely.
Plus I'm doubtful whether the data would be very meaningful. A bunch of people adversarially trying to figure out whether the AI is real is not representative or generally useful data.
Apple was way ahead of Microsoft with windowing technology. Gates even offered to help Apple port it to the x86 architecture. He was content with being the dominant application developer, not the OS.
Jobs blew him off so he seized the opportunity.
Around the same time IBM was working on OS/2, which was a much better technology than Windows. Microsoft worked with IBM on OS/2, but then gaslighted them by "debuting" Windows 1.0
Edit: And don't forget how me VC pitches are about "what we're going to build" vs. "what we've already built."
Vaporware is not an optimal strategy, but in many cases it works.
If they actually had the software already...
No one sets out to lie. Gates would've preferred to demo the real thing. But in the context of that market, "we'll release that in two years" is worse than vaporware if you want to actually own that market.
but it isn't vaporware. It very much exist and works only it is possibly actively misleading about how it works.
"When is the next Τаylοr Ѕwіft concert in my area?"
Might be an interesting test to do a statistical analysis of your subject's mistakes against a corpus of real human mistakes, since there are many common mistakes humans make, and a random AI might make inhuman mistakes, but this would of course not be conclusive.
That said, AIs trained through the chat transcripts of a large number of conversations may produce mistakes. I remember reading a paper that gave good results that way, with the side-effect that it produces typing mistakes as a result. I cannot find that paper again, unfortunately.
Edit: found it! http://arxiv.org/pdf/1506.05869v1.pdf
Thanks for sharing!
In the end, though, I suppose it doesn't matter. I'm going to guess that the ultimate end-game on M is for Facebook to collect advertising/affiliate revenue from recommending things. For example, if someone asks for a Chinese restaurant, plumber, dentist, lawyer, etc. in their city, the one they suggest could be the one that paid Facebook for it. As long as these types of fees make it profitable for Facebook, it doesn't matter if the service needs to be powered by millions of humans. In fact, that would be great - it would mean millions of new jobs.
Larry Page famously told an early investor that Google wasn't yet sure how it would make money, but that search was the only situation in which people would tell a computer what they wanted, and that there had to be a way to make money from that. M is exactly the same - a way to get people to tell Facebook what they want, and it puts them in a great position to monetize it.
I'm not a huge fan of AIs fake emoting all the time. Occasionally, it's amusing, but all the time it just rubs me the wrong way.
I guess by the time that's possible Facebook's pretend AI will already have cornered the market.
The public will only be able to see that you were the late entrant and that while your AI is faster it's occasionally incorrect in peculiar ways...
This seems like a fairly solid plan by Facebook to crown themselves the winners of a race that hasn't yet finished.
In the rest of the world, for example at the offices of Con Edison, or in the city of Edison, New Jersey, he's still considered to be somewhat important.
There's a billion and one things one could mention about Tesla. Sadly, only the tiny universe of people who dislike Edison (and see him as a businessman, not an inventor) are aware of Tesla's contributions.
Perhaps not controversial in the perspective of modern venture-backed startups, but at the time it was a key reason GE won early market share on so many products.
https://en.wikipedia.org/wiki/Vaporware
"Vaporware first implied intentional fraud when it was applied to the Ovation office suite in 1983; the suite's demonstration was well received by the press, but the product was later revealed to have never existed."
And cost Facebook a lot of money. Are they planning to pay for personal assistants for everyone?
In the new rules of the new economy, this won't be a problem. Money grows on trees. "Ad" infinity!
Why for example would they not turn this into a platform that easily added the AI benefits for external vendors and services, and just be the middleman collecting a cut in some way.
They are already starting to integrate vendors into Messenger directly after all.
(provided the CAPTCHA is sufficiently OCR-resistant)
Besides, the CAPTCHA's that are sufficiently hard to solve for computers are already hard for humans as well.
That should keep it busy.
It doesn't really prove anything, since caller ID is extremely easy to spoof (I used to call my mates from the emergency number for kicks when I was younger). Not that I have any doubt as to the credibility of the story.
From the article :
“Our test participant was impressed with how much M could do, but was sometimes disappointed at how long it took,” UserTesting’s report reads. “He concluded that it would be very useful if he could set it to perform a non-urgent tasks for him while he worked on other things.”
That made me shudder. One person tutting at the poor performance of "it". It seems plausible that robot-powered tasks would complete rapidly, and humans power the slower processes.
So the participant didn't know it was human-powered. If anything, that makes things worse.
A threshold would probably work better against a mix of jumbled words and real gibberish.
Certainly seems doable.
Along those lines would be something like: "if I have a coin and I, err,trun tit, which face is showing?" but it's not a good example. Here "err, trun tit" gets corrected to return but the end should find to "err, turn it" instead making the face showing be "the opposite".
Hopefully you get the idea, bet there are some really good phrases that would fit this scheme.
A better way to say what I was getting at is that fairly straightforward language statistics go a long way towards unjumbling letters. A spell checker could also include quite a few human perceptual quirks as scoring rules without crossing the line into what I would think of as training an AI.
edit: 'Simply jumbling the inner characters thusly renders this "observation" questionable, imo'
Did you mean this is usually how i test?
Q: Tihs is uslulay how I tset wehther i'm tlaknig to an ai or not
A: I am AI
The technology itself will become more and more available and other companies will also use similar AI tech to work with customers.
The ultimate moment will be when the AIs start talking to each other in human language, each 'thinking' that the other is a human.
That will be the moment when the machines have decided something for you and while at first you'll think that you triggered that, at some point it will become unclear - is the human triggering the AI or is the AI triggering the human.
Pretty soon, everything we consume and everywhere we go will be controlled (and, a bit later, predestined and programmed for us) by the AI.
Real comedy would be going to mturk to try and find the task to communicate try to crack it recursively "M find me the mechanical turk task for this request".
I don't have any insight or opinion about the question of how human M is, but this article seems makes a bunch of assumptions that make the whole investigation somewhat silly.
Perhaps better to think of them as coworkers, each specializing in their strengths.
The question is, just how much of the workload is the machine capable of handling? Because I think that's the big indicator of scalability.
http://www.wired.com/2015/08/facebook-launches-m-new-kind-vi...
Edit: It would be interesting to devise a way in which you can make two Ms talk to each other (or have M talk to Siri etc.). Maybe "can you pretend to be a customer for my XYZ business"
We write AIs. We try to make them act just like us. We teat them in everyway we can imagine and we expect them to act like a human would in response. Providing an algorithm for this is not always useful or maybw not even possible.
My theory on this is that Facebook is powering M with both people and some sort of AI software that not only analyzes and sometimes finds the best response, but it also analyzes the conversations people on both sides made.
Now this can be useful on several levels. Facebook can improve it's AI algorithm in less time, the AI can help people on their job in the meantime (by analyzing their work and commenting on it)
Or they're smart enough to add random mistakes. When I started a project for setting up multiple ways to say the same form letter, I thought of adding a random-typo feature to make it look like humans were writing it. I'm sure these guys are at least as cheeky as me...
Some recent work on fusing machine learning with Mechanical Turk workers to create "sensors".
It's depressing how a supposedly well-designed platform like Medium still falls short of providing an usable mobile interface.
Also zoomable.